A Computer Vision Application for Vehicle Collision and Damage Detection

During the car insurance claims process after an accident has occurred, a vehicle must be assessed by a claims adjuster manually. This process will take time and often results in inaccuracies between what a customer is paid and what the damages actually cost. Separately, companies like KBB and Carfax rely on previous claims records or untrustworthy user input to determine a car’s damage and valuation. Part of this process can be automated to determine where exterior vehicle damage exists on a vehicle. 

In this project, a deep-learning approach is taken using the MaskR-CNN model to train on a dataset for instance segmentation. The model can then outline and label instances on images where vehicles have dents, scratches, cracks, broken glass, broken lamps, and flat tires. The results have shown that broken glass, flat tires, and broken lamps are much easier to locate than the remaining categories, which tend to be smaller in size. These predictions have an end goal of being used as an input for damage cost prediction.